feat(CalibrationClassification): add PTQ int8 optimization and evaluation for onnx and tensorRT

[vividf]

Summary

This PR:

• Introduces post-training quantization (PTQ) with INT8 precision for the CalibrationClassification component.
• Adds support and evaluation scripts for both ONNX and TensorRT inference using INT8.

For easier review, the file name will remain visualize_lidar_camera_projection.py in this PR. It will be renamed to toolkit.py in a follow-up change.

Model Performance Comparison: FP16 vs INT8

Metric	FP16	INT8
Total Samples	6,800	6,800
Correct Predictions	6,336	6,319
Accuracy	93.18%	92.93%
Average Latency	5.29 ms	2.46 ms ✅

Observation:

• INT8 is ~2.15× faster than FP16 (lower latency).
• Accuracy drop is minimal (0.25 pp).

Per-Class Accuracy

Class	Description	FP16 Accuracy	INT8 Accuracy
0	Miscalibrated	99.21%	99.53% ✅
1	Calibrated	87.15%	86.32% 🔻

Observation:

• INT8 slightly improves miscalibrated class detection.
• Small drop (~0.8 pp) in calibrated class accuracy.

Change point

Explain for detail

Note

Test performed

• Log

Log output

[vividf]

Will separate the PR